Abstract
Detecting microseismicity is a crucial processing step for monitoring well injection operations. In recent years, distributed acoustic sensing (DAS) technologies have gained popularity for this purpose due to their high receiver density and cost-effectiveness. However, DAS recordings typically have lower signal-to-noise ratio (S/N) and generate significantly larger amounts of data compared to classic borehole geophones. This poses challenges for conventional signal processing methods. We present a versatile method for detecting microseismic events in continuous DAS recordings using a binary semantic image segmentation model. This approach assigns each pixel in a DAS input image its probability of being part of a seismic event, separating the recorded wavefield from background noise. To train the model, we create a case-study-specific synthetic data set that includes a wide range of wavefield phenomena. Labels are generated automatically based on amplitude exceedances of the synthetic strain responses before being inserted into ambient noise samples. We apply this method across three stages of hydraulic fracturing monitored with two differently oriented fibers. Although the model was trained exclusively on synthetic data, it reliably identifies microseismic events in the field data. Furthermore, high-quality predictions of the model hold the potential to be used for arrival-time picking.